About this Research Topic

Manuscript Submission Deadline 04 March 2022
Manuscript Extension Submission Deadline 05 September 2022

The problem of continual learning has recently been the object of much attention in the machine learning community, yet this has mainly been approached from the point of view of preventing the model being updated in the light of new data and ‘catastrophically forgetting’ its initial, useful knowledge and abilities. A typical example is that of an object detector which needs to be extended to include classes not originally in its list (e.g., ‘donkey’ in a farm setting), while retaining its ability to correctly detect, say, a ‘horse’. The unspoken assumption being that we are quite satisfied with the model we have, but wish to extend its capabilities to new settings and classes.

Much recent work in continual learning for computer vision has focused on class-incremental learning, under the assumption that complete supervision is available. This way of posing the problem of continual learning is, however, in rather stark contrast with common real-world situations in which an initial model is trained using limited data, only for it to then be deployed without any additional supervision.

This Research Topic aims to explore, and highlight to the wider computer vision community, a new continual learning setting in which models are continually updated using unsupervised data that arrives either as a stream, periodically, or as occasional chunks – for example, a person detector used for traffic safety purposes on a busy street. Even after a system has been trained extensively on the many available public datasets, experience shows that its performance in its target setting will likely be less than optimal. Nevertheless, such a model can learn from the data it continuously acquires, not to learn new classes but to better perform its task.

A significant feature of such real-world scenarios is that the process that generates the data varies with time (as in the night/day and weekly/annual cycles in the data captured by a camera outside an office block entrance). This requires approaches to move well beyond the naive concept of ‘tasks’, towards modeling and leveraging the dynamics of the data.

In order to tackle these challenges, it is crucial that suitable realistic benchmark datasets be introduced (for instance in activity recognition or object detection), to enable the field to mature and abandon existing toy datasets in which the ordering of the data is arbitrarily set, rather than being natural.

Topics of interest include, but are not limited to:

- Analyses of the suitability of existing datasets for continual learning in computer vision
- New benchmark datasets explicitly designed for continual learning in computer vision, in particular in the unsupervised setting
- Protocols for training and testing in unsupervised continual learning
- Metrics for assessing unsupervised continual learning
- Relationships between continual learning and online learning
- Computer vision applications of continual learning
- Applications of transfer learning, multi-task and meta-learning for unsupervised continual learning
- Relationships between unsupervised continual learning, life-long learning and few-shot learning

Keywords: Continual Learning, Life-long Learning, Online Learning, Computer Vision, Time Series


Important Note: All contributions to this Research Topic must be within the scope of the section and journal to which they are submitted, as defined in their mission statements. Frontiers reserves the right to guide an out-of-scope manuscript to a more suitable section or journal at any stage of peer review.

The problem of continual learning has recently been the object of much attention in the machine learning community, yet this has mainly been approached from the point of view of preventing the model being updated in the light of new data and ‘catastrophically forgetting’ its initial, useful knowledge and abilities. A typical example is that of an object detector which needs to be extended to include classes not originally in its list (e.g., ‘donkey’ in a farm setting), while retaining its ability to correctly detect, say, a ‘horse’. The unspoken assumption being that we are quite satisfied with the model we have, but wish to extend its capabilities to new settings and classes.

Much recent work in continual learning for computer vision has focused on class-incremental learning, under the assumption that complete supervision is available. This way of posing the problem of continual learning is, however, in rather stark contrast with common real-world situations in which an initial model is trained using limited data, only for it to then be deployed without any additional supervision.

This Research Topic aims to explore, and highlight to the wider computer vision community, a new continual learning setting in which models are continually updated using unsupervised data that arrives either as a stream, periodically, or as occasional chunks – for example, a person detector used for traffic safety purposes on a busy street. Even after a system has been trained extensively on the many available public datasets, experience shows that its performance in its target setting will likely be less than optimal. Nevertheless, such a model can learn from the data it continuously acquires, not to learn new classes but to better perform its task.

A significant feature of such real-world scenarios is that the process that generates the data varies with time (as in the night/day and weekly/annual cycles in the data captured by a camera outside an office block entrance). This requires approaches to move well beyond the naive concept of ‘tasks’, towards modeling and leveraging the dynamics of the data.

In order to tackle these challenges, it is crucial that suitable realistic benchmark datasets be introduced (for instance in activity recognition or object detection), to enable the field to mature and abandon existing toy datasets in which the ordering of the data is arbitrarily set, rather than being natural.

Topics of interest include, but are not limited to:

- Analyses of the suitability of existing datasets for continual learning in computer vision
- New benchmark datasets explicitly designed for continual learning in computer vision, in particular in the unsupervised setting
- Protocols for training and testing in unsupervised continual learning
- Metrics for assessing unsupervised continual learning
- Relationships between continual learning and online learning
- Computer vision applications of continual learning
- Applications of transfer learning, multi-task and meta-learning for unsupervised continual learning
- Relationships between unsupervised continual learning, life-long learning and few-shot learning

Keywords: Continual Learning, Life-long Learning, Online Learning, Computer Vision, Time Series


Important Note: All contributions to this Research Topic must be within the scope of the section and journal to which they are submitted, as defined in their mission statements. Frontiers reserves the right to guide an out-of-scope manuscript to a more suitable section or journal at any stage of peer review.

Topic Editors

Loading..

Topic Coordinators

Loading..

articles

Sort by:

Loading..

authors

Loading..

views

total views article views article downloads topic views

}
 
Top countries
Top referring sites
Loading..

Share on

About Frontiers Research Topics

With their unique mixes of varied contributions from Original Research to Review Articles, Research Topics unify the most influential researchers, the latest key findings and historical advances in a hot research area! Find out more on how to host your own Frontiers Research Topic or contribute to one as an author.